The Compute Bill Always Comes
Kemdirịm@kemdirim
−12.1%
since Sep 30
The idea
I think the AI coding tools themselves are interchangeable, but every session ends as a bill paid to whoever sells the underlying compute. So this basket skips the app layer and owns the stack that collects regardless of which tool or lab wins: chipmakers at the core, memory and networking, fab equipment makers, GPU capacity renters, and the power and data center layer feeding it all. About 10% sits in cash to add on pullbacks, and the thesis holds as long as AI workloads keep needing more compute, not less.
What it owns
- Every AI model running on silicon means a bill, and these companies design or manufacture the chips that process nearly all of it. · NVIDIAIt manufactures the GPUs that process most AI workloads, so the majority of the compute bill flows through its chips., Taiwan SemiconductorIt manufactures AI chips designed by NVIDIA, AMD, and hyperscalers alike, collecting on the compute bill across all chip makers., BroadcomIt designs custom chips for hyperscalers building their own AI silicon to reduce their reliance on NVIDIA., Advanced Micro DevicesIt manufactures AI chips as the primary alternative to NVIDIA, so it wins if the compute bill spreads across multiple chip suppliers., Marvell TechnologyIt designs custom silicon and connection chips deployed inside AI data centers, another line item on the compute bill.30%
- GPUs need fast memory and high-speed connections to work together, and both components are scarce relative to demand. · Micron TechnologyIt manufactures high-bandwidth memory that must be paired with every AI chip, and only three companies make it., Credo TechnologyIt makes the high-speed interconnect chips and cables that link GPU clusters together, a requirement that scales with each new installation., Arista NetworksIt manufactures networking switches that connect GPU clusters together, a component required in every large AI buildout., Astera LabsIt manufactures interconnect components that let AI chips communicate with each other and with memory, designed into every new AI server.17%
- Building more AI chips requires new factories, and only a few companies make the specialized machines that go inside them. · ASML HoldingIt makes the lithography machines required to manufacture advanced AI chips, and no competitor can currently replicate this technology., Lam ResearchIt manufactures etching tools used in semiconductor fabs to produce AI chips and memory, collecting on every expansion of manufacturing capacity., Applied MaterialsIt manufactures a broad range of tools used in semiconductor fabs, making it a diversified play on chip production expansion., KLA CorporationIt manufactures inspection equipment used to detect defects in advanced chips, and more complex AI chips require more inspections per wafer.16%
- Computing at scale consumes enormous electricity, and the suppliers of power infrastructure and cooling collect a bill regardless of which model wins. · Vertiv HoldingsIt supplies the cooling and power distribution equipment that keeps AI data centers operating, a cost that applies regardless of which model wins., Constellation EnergyIt provides nuclear power through long-term contracts with hyperscalers and data centers expanding their AI infrastructure., EatonIt manufactures electrical equipment that distributes power from the grid into data center racks, a bill that grows with every new facility.14%
- Most labs and software tools lease GPU capacity instead of owning it, and these companies operate the data centers that supply it. · CoreWeaveIt rents GPU capacity to customers who need compute power, collecting directly from the growing shortage of available GPUs., Nebius GroupIt rents GPU capacity to customers seeking alternatives to larger providers, growing by capturing share of the compute shortage., OracleIt operates data centers for AI training and inference, collecting revenue from multi-year contracts with hyperscalers and labs.13%
- Cash10%